To the point of the training data, I just want to add that copyrighted works have never prevented people from viewing and studying the art, and even emulating the styles. I would say that every artist is standing on the shoulders of those that came before them.

I personally see the training data in the same way. The AI is learning from all of it, and using a complex, multi-dimensional latent space to "understand" how it all relates and then making new outputs from that understanding, like an artist would after decades of exposure and practice.

Copyrights protect an artist's piece of work, not their style. In the aforementioned 5,000 generations, I have never once "recreated" a famous work.